Testing for unit roots with stationary covariates

Transcription

1 Journal of Econometrics 115 (2003) Testing for unit roots with stationary covariates Graham Elliott a;, Michael Jansson b a Department of Economics, University of California, San Diego, Gilman Drive, La Jolla, CA 92093, USA b Department of Economics, University of California, Berkeley, 549 Evans Hall #3880, Berkeley, CA , USA Accepted 12 January 2003 Abstract We derive the family of tests for a unit root with maximal power against a point alternative when an arbitrary number of stationary covariates are modeled with the potentially integrated series. We show that very large power gains are available when such covariates are available. We then derive tests which are simple to construct (involving the running of vector autoregressions) and achieve at a point the power envelopes derived under very general conditions. These tests have excellent properties in small samples. We also show that these are obvious and internally consistent tests to run when identifying structural VARs using long run restrictions. c 2003 Elsevier Science B.V. All rights reserved. JEL classication: C3 Keywords: Unit roots; Power envelopes; Structural vector autoregressions 1. Introduction Due to the eects of the assumption of a unit root in a variable on both the econometric method used and the economic interpretation of the model examined, it is quite common to pre-test the data for unit roots. This is typically done by either (or both) testing variables one by one for unit roots or by examining cointegrating rank using Johansen (1988) tests or their asymptotic equivalent. In testing variables one by one, commonly the t-test method of Dickey and Fuller (1979) is employed. This hypothesis test is asymptotically optimal when the data are stationary and is a natural statistic to consider. However, in the unit root case there Corresponding author. Tel.: ; fax: addresses: gelliott@weber.ucsd.edu (G. Elliott), mjansson@econ.berkeley.edu (M. Jansson) /03/$ - see front matter c 2003 Elsevier Science B.V. All rights reserved. doi: /s (03)

2 76 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) are many other tests available that have greater power. Elliott et al. (1996) (denoted ERS in the remainder of the paper) showed that there is no uniformly most powerful test for this problem and derived tests that were approximately most powerful in the sense that they have asymptotic power close to the envelope of most powerful tests for this problem. This paper considers a model where there is one series that potentially has a unit root, and that this series potentially covariates with some available stationary variables. In a model similar to the one examined here, Hansen (1995) demonstrated in a model with no deterministic terms that no uniformly most powerful test for a unit root in the presence of stationary covariates exists and that power gains are to be had from using these covariates. He suggested covariate augmented Dickey Fuller (CADF) tests and showed that these tests had greater power than tests that ignored these covariates. 1 This paper extends the results in Hansen (1995) in a number of ways. First, we show that the point optimal tests implicit in the power envelope derived in Hansen (1995) and computed when all nuisance parameters are known are feasible when these parameters are not known. We also extend the results by deriving the power envelope in the more empirically relevant cases of where constants and/or time trends are also included in the regression. We propose tests that are feasible to construct with data and attain the power envelope at a point. These tests have good power at other points as well. We then show that these are natural tests to report in justifying the unit root assumption in the popular method of identifying structural vector autoregressions (VAR s) from long run restrictions (as suggested by Blanchard and Quah, 1989). The paper is set up as follows. In the next section the model is introduced, and the power bounds for the problem are established. In the third section, tests which feasibly attain these power bounds at a point are derived and discussed. Section 4 examines the tests empirically using Monte Carlo methods. A fth section discusses the tests as they relate to identifying structural VAR s from long run restrictions. The nal section concludes. All proofs are contained in a separate appendix, available from the authors upon request. 2. Model and power envelopes Consider the model z t = t + u t ; t =1;:::;T (1) and ( ) (1 L)uy;t A(L) = e t ; (2) u x;t where z t =[y t ;x t] ;x t is an m 1 vector, y t is 1 1, 0 =[ y0 ; x0 ], 1 =[ y1 ; x1 ], u t =[u y;t ;u x;t] and A(L) is a matrix polynomial of nite order k in the lag operator L. 1 There is also a discussion of this work in Caporale and Pittis (1999).

3 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) For the constructed test statistics we will assume that (A1) A(z) = 0 has roots outside the unit circle. (A2) E t 1 (e t )=0; E t 1 (e t e t)= and sup t E e t 2+ (a.s.) for some 0, where is positive denite and E t 1 (:) denotes conditional expectation with respect to {e t 1 ;e t 2 ;:::}. (A3) u 0 ;u 1 ;:::;u k are O p (1). Dene u t ()= (1 L)u y;t u x;t with spectral density at frequency zero (scaled by 2), so we have = A(1) 1 A(1) 1 where we can partition this after the rst column and row so that [ ]!yy! yx =! yx xx (we partition similarly). We will further dene R 2 =! 1 yy! yx 1 xx! yx, the frequency zero correlation between the shocks to x t and the quasi-dierences of y t.ther 2 value represents the contribution of the stationary variables it is equal to zero when there is no long run correlation and one if there is perfect correlation. We impose that R 2 1, hereby ruling out the case under the null where the partial sums of x t cointegrate with y t. If there is such a cointegrating relation, this should be modeled in the system taking the model outside this framework (unless the coecients of the cointegrating vector are known, in which case the model can be rotated back into this framework, see Elliott et al., 2002). We consider ve cases indexed by superscript i (i =1; 2; 3; 4; 5) for the deterministic part of the model (where parameters are free unless otherwise stated): Case 1: y0 = y1 = 0 and x0 = x1 =0. Case 2: y1 = 0 and x0 = x1 =0. Case 3: y1 = 0 and x1 =0. Case 4: x1 =0. Case 5: No restrictions. Each of these cases can be characterized by the restriction (I 2(m+1) ( S i ) =0 ) 1 0 where =[ 0 1 ] ; S i isa2(m +1) 2(m + 1) matrix where S 1 =0, S 2 =, 0 0 ( ) ( ) Im+1 0 Im+2 0 S 3 =, S 4 = and S 5 is the identity matrix This represents a fairly general set of models in which we have a VAR in the model of x and the quasi-dierence of y. We wish to test that the parameter is equal to one (y t has a unit root) against alternatives that this root is less than one. Following the general methods of King (1980, 1988) we will examine Neyman Pearson tests for this hypothesis. Following the application of these methods to testing for unit roots in ERS and Elliott (1999) we will examine Neyman Pearson tests for this hypothesis under simplifying assumptions, and then in the following section we will derive general tests that are asymptotically equivalent to these optimal tests.

4 78 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) With the assumption that A(L)=I (so that = ) and assuming the e t are normally distributed and u y0 =0, we will examine tests against the local alternative that c=c 0 where =1+c=T and =1+ c=t with c; c being xed (we will suppress the dependence of on T in the notation). The system likelihood ratio test statistic for the hypothesis is given by T T i (1; )= û i t() 1 û i t() û i t(1) 1 û i t(1); t=1 where we have for r =; 1 that û i t(r)=z t (r) d t (r) ˆi (r); t=1 where z t (r) = [(1 rl)y t ;x t] for t 1 and z 1 (r)=[y 1 ;x 1 ], [ ] 1 r 0 (1 rl)t 0 d t (r) = for t 1; 0 I m 0 I m t [ ] d 1 (r) = ; 0 I m 0 I m and [ ( T ) ] [ ] T ˆ i (r)= S i d t (r) 1 d t (r) S i S i d t (r) 1 z t (r) ; t=1 t=1 where D is the Moore Penrose inverse of D. The test has rejection regions of the form {y t ;x t : i (1; ) c b} where b is a critical value. Theorem 1. 2 For the model in (1) and (2) with A(L) =I, e t independent N (0;) random variables and (A3) holding then with =1+c=T and =1+ c=t with c, c xed as T, the most powerful test of H 0 : c=0 vs: H a : c=c 0 has asymptotic power functions P(c; c; R 2 )=Pr[ i (c; c; R 2 ) b(c; R 2 )]; where i (c; c; R 2 )=g i (c; c) +(c 2 2cc)Q (W1c i )2 +2cQ 1=2 W1c i dw 2 + h i (c; c; R 2 ); b(c; R 2 ) is a constant, Q = R 2 =(1 R 2 ); W1c 1 = W 1c;W1c i = W 1c W 1c for i =2; 3; 4 and W1c 5 = W 1c (4 6s) W 1c (12s 6) sw 1c ;g i (c; c)= c 2 W1c 2 cw 1c(1) 2 for i =1; 2; 3 and g i (c; c)= c 2 W1c 2 +(1 c)w 1c(1) 2 k 1 [(1 c)w 1c (1) + c 2 sw 1c ] 2 for i =4; 5; h i (c; c; R 2 ) is zero except for h 4 (c; c; R 2 )=k 1 [(1 c)w 1c (1) + c 2 sw 1c ] 2 (k + Q c 2 =12) 1 {(1 c)w 1c (1) + c 2 sw 1c + Q[ c 2 (c c) W 1c c(c c) sw 1c ]+ Q 1=2 [c s dw 2 c 2 dw2 ]} 2, and k =1 c +c 2 =3. All integrals are 0 to 1 over s with s suppressed, so e.g. W 1c = l 0 W 1c(s)ds and W 1c (s)=c s 0 ec(s ) W 1 ()d + W 1 (s), W 1 and W 2 are independent univariate standard Brownian motions. 2 Proofs are available in a UCSD discussion paper version of this paper.

5 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) In case 1 this is apart from a scale factor the same as that reported in Hansen (1995). 3 A number of features are worth noting. Firstly, the dependence of the test on c indicates that no uniformly most powerful test is available for this problem, power depends on the choice of the alternative. Second, the distribution of the test is nonstandard. Third, the optimal test statistic depends on and its distribution depends on the parameter R 2. When R 2 = 0 then i (c; c; R 2 )=g i (c; c) which is equivalent to the asymptotic limits of the tests derived in ERS, thus the most powerful tests coincide asymptotically with tests with the relevant invariance properties with respect to deterministic terms which do not use the information in the covariates (cases 1 3 equate to the constant included case, cases 4 5 are the time trend included result). When R 2 is nonzero the optimal univariate and system tests are dierent, indicating that information is lost when information in the covariates is ignored. The results in the Theorem give the local power for any choice of c at any local alternative c. When we set c=c, we obtain by construction the test that has the highest attainable power. By evaluating the powers setting c = c we obtain the envelope of greatest asymptotic power, which we call the power envelope. Fig. 1 examines the power envelopes for various R 2. The power envelope when R 2 = 0 has the lowest power this is the relevant envelope if no covariate information is employed. When R 2 is greater than zero, the power attainable increases considerably above this lower bound. Hence, use of covariates has the potential to greatly increase the power of tests for a unit root, as indicated by Hansen (1995). The larger the R 2, the more powerful the optimal test. 4 These results are true for each of the various assumptions on the deterministic terms. 5 Comparing the rst two panels in Fig. 1 we see the eect of estimating the constant terms. This eect is small, e.g. when R 2 =0:5 and c = 5 the power envelope in the constants known case is 70% whilst when the constants are unknown this power is 62%. Both of these powers are substantially above that of the case where no covariates are employed, where the envelope attains a power of 32%. As in the case where there are no covariates, the eect on the power envelopes for the case where the trend terms (coecients on time trends) are not known is quite large. In the case mentioned above, where R 2 =0:5 and c = 5 the maximal power in case 5 is 33%, far below the 62% when only coecients on the constants are known. Note though that the maximal power in this case even when constants and coecients on the time trend are estimated is (just) above that for the case where stationary covariates are ignored and the coecient on the time trend is known. In general, the power losses from not knowing the coecient on the trends in the x t regressions is small (dierences between cases 4 and 5, not pictured in the gures), between zero (when R 2 is small) and 6% or so (when R 2 is large). There is clearly the 3 We also have a notational dierence in that our R 2 is dened in Hansen (1995) as 1 R 2. We changed the notation to accord with the usual use of R 2. 4 The asymptotic results are not appropriate at R 2 = 1, which is readily seen from the limit expression which would not be nite at this point. 5 Cases 2 and 1 are asymptotically identical, so we omit case 2. Case 4 has functions similar to case 5 and is omitted.

6 80 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Fig. 1. Power envelopes for cases 1, 3 and 5, respectively. Note: Envelopes for R 2 = 0, 0.3, 0.5, 0.7 and 0.9 where power is increasing in R 2. potential for much to be gained in terms of power from exploiting stationary covariates in constructing tests for a unit root. The construction of tests that achieve these gains is addressed in the next section.

7 3. Feasible tests G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) In this section we derive families of tests that asymptotically attain the power bounds derived above at pre-specied points, relaxing the normality and known nuisance parameter assumptions. The method for constructing the test is set out in four steps (a) Estimate nuisance parameters for detrending and R 2. Run a VAR A(L)z t (1) = deterministics + e t including no deterministic terms for case 1, constants for cases 2 and 3, constants and time trends for cases 4 and 5. Using the residuals from the VAR 6 construct ˆ = T 1 T t=k+1 ê tê t; ˆ = Â(1) 1 ˆÂ(1) 1 and ˆR 2 = ˆ! yx ˆ 1 xx ˆ! yx = ˆ! yy where Â(1) = I + k i=1 A i and A i is the (i + 1)th matrix element of A(L). (b) Construct detrended data under the null and alternative hypotheses, i.e. construct for r =(1;), ũ i t(r)=z t (r) d t (r) i (r); where [ ( T ) ] [ ] T i (r)= S i d t (r) ˆ 1 d t (r) S i S i d t (r) ˆ 1 z t (r) : (3) t=1 t=1 (c) Run VAR s (for r =1; ), i.e. run Ã(L)ũ i t(r)=ẽ t (r) and construct the estimated variance covariance matrices T (r)=t 1 ẽ t (r)ẽ t (r) : t=k+1 (d) Construct the test statistic i (1; )=T (tr[ (1) 1 ()] (m +)): This test will have asymptotic power that achieves the power bound at c under the assumptions. Theorem 2. For the model in (1) and (2) with assumptions (A1) (A3) holding and deterministic terms correctly specied for each case then as T i (1; ) i (c; c; R 2 ); where denotes weak convergence. Thus the critical values for the test depend on the alternatively chosen ( c) and R 2. The feasible test asymptotically achieves the highest power possible at c. We have chosen here to let c = 7 for cases 1 3 and c = 13:5 for cases 4 and 5 (which follows the choice of ERS). In principle and practice we could choose dierent values 6 In practice one can choose the lag length of the VAR through theory or a consistent lag length estimator such as the BIC information criterion.

8 82 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Table 1 Asymptotic critical values (distribution in Theorem 3) R Cases 1, Case Case Case Note: Critical values were computed using 1500 steps as approximations to the Brownian motion terms in the limit theorem representations and replications. The critical values reported are for tests of size 5% with c = 7 for cases 1, 2 and 3 and c = 13:5 for cases 4 and 5. for c depending on R 2 ; however, as R 2 rises above zero lack of power becomes less problematic so it seems reasonable to us to choose c for the worst case scenario. Asymptotic critical values for the test for selected values of R 2 are given in Table 1. The relevant critical value is determined for the estimated ˆR 2. For values of R 2 between the ones given in Table 1, interpolation can be used to approximate the critical value. 4. Evaluation of the tests 4.1. Large sample evaluation Fig. 2 examines the power of the feasible test for cases 1, 3 and 5 in each panel, respectively. The gures give the results for R 2 =0:3; 0.5 and 0.7. Accompanying the power curves are the power envelopes for comparison. The feasible point optimal test has power that is close to the power envelope, suggesting that there is little asymptotic power loss at points away from where the test is optimal, especially for lower values for R 2. This is similar to results of ERS, where for R 2 = 0 this was found to be true. When R 2 =0:5 the dierence between the power envelope and the asymptotic power of the feasible test is small for alternatives at moderate distance and further from the null, but a little larger for alternatives close to the null. This becomes more apparent for larger R 2. To the extent that very large values for R 2 are probably not too relevant empirically, this may not be too much of a problem. The suggestion from these graphs appears to be that the most useful choice of c in practice may depend on R 2. We also examined the power curves for the case where c= 7 perhaps improve the closeness of the power curves to the envelopes for these near alternatives. When this alternative is chosen this indeed happens; however, the tradeo is that the power curves for R 2 small are not as close to the envelope for more distant alternatives. Thus we recommend that choosing c = 13:5 as power is more of a concern when R 2 is small. The power gains are clearly substantial for each of the cases for the deterministic terms (results for case 4 are similar to those for case 5). Consider the gains from using covariates when R 2 =0:5. At the local alternative c= 5, in case 3 power rises by 30% and in case 5 by 35%. Such gains in power substantially improve the odds of correctly distinguishing a process with a unit root from a slowly mean reverting process.

9 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Fig. 2. Power envelopes and power curves for cases 1, 3 and 5. Note: Unbroken lines are envelopes for R 2 =0:3, 0.5 and 0.7 and broken lines are power of point optimal tests for each R 2 where power is increasing in R 2 : 4.2. Small sample evaluation We will examine various special case models in samples of 100 observations. Along with the above tests, we report results for the commonly applied test of Dickey and Fuller (1979) and also the P T test of ERS as well as the Hansen (1995) CADF test.

10 84 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Table 2 Small sample results for i (1; ) DF PT i (1; ) R Case 1: No deterministic terms Case 3: Constants in each regression Case 5: Constants and time trends in each regression Note: Based on replications of the model with T = 100, normal errors as discussed in the text. The system test is implemented with R 2 estimated. Table 2 reports results of simulations of the model in (1) and (2) for each of the cases 1, 3 and 5, respectively, where A(L) =I (and this is known), e t is normally distributed with variances equal to 1 and R 2 as reported in the table. Size is given in the row corresponding to = 1 and (empirical) power against the indicated alternatives in the following rows. When there are no deterministic terms in the model the DF and P T single equation tests do similarly well (see ERS for a discussion of this similarity). In the test proposed here, when R 2 =0 power and size are comparable to the univariate tests indicating that even in small samples little may be lost by including extraneous information and doing the system test. As R 2 increases, size remains well controlled whilst power rises considerably. Consider the case of the true being equal to 0.96, the P T test has power around 23% whilst if R 2 =0:25 the system test has power equal to 34%, roughly a 50% gain.

11 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Table 3 Eect of estimating R 2 R 2 known Estimated R 2 R Case 3: Constants in each equation Case 5: Constants and time trends in each equation Note: As per Table 2. When a constant is included, the P T statistic gains in power over the Dickey and Fuller (1979) t test are very large. Again, when R 2 = 0 the test proposed here has similar size and power to the P T statistic indicating that little is lost adding extraneous stationary covariates. In general, size is less well controlled, especially for R 2 close to one (where the asymptotic theory would no longer be relevant; however, it would not be expected that such models would be appropriate for real world data). There is some evidence of power losses from not knowing the constant term. At a value of =0:96 the power when the constant is known (or zero) is 49% compared to the unknown constant power of 45% when R 2 =0:49. Even so, power for the test with the constant unknown is quite high in many cases, and is far beyond that achievable when covariates are not employed. Similar results are found for the detrended (case 5) model. In both of these cases power when using covariates is substantially greater than when relevant covariates are ignored (for example, in case 3 when =0.9, power of the test proposed here when R 2 =0.25 is 20% for the Dickey and Fuller test and is 49% for the test with covariates employed. There are as usual power losses in including a time trend. In the case of =0.96 and R 2 =0.25 the power drops from 36% in case 3 to 13% in case 5. The eect of estimating R 2 in the computation of the test is examined in Table 3 (for cases 3 and 5 in each of the panels, respectively). Here the results when R 2 is estimated are repeated from Table 2 on the right-hand side panels, whilst the same results using the critical value chosen using the true R 2 are given in the left-hand panels. There is very little dierence, even in a sample of 100 observations. Most of the dierences in size and power are at the third decimal place. It is only for case 5

12 86 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Table 4 CADF and i (1; ) CADF i (1; ) R Case 3: Constants in each equation Case 5: Constants and time trends in each equation Note: As per Table 3. The CADF refers to the test procedure in Hansen (1995). In each case the same R 2 estimate is used to determine the critical value. when R 2 is a little larger that there is much of an eect, but the eect is minor (in these cases there is a small power loss from estimating R 2 ). Table 4 compares the CADF test of Hansen (1995) with the feasible test derived here (again for the leading cases 3 and 5, respectively). The CADF test augments the usual Dickey and Fuller (1979) test with lags, leads and the contemporaneous values of x t. In this table, with no serial correlation, this amounts to including x t as a regressor in the ADF regression and then constructing the t-test of the unit root hypothesis as normal. As shown in Hansen (1995) this test also depends on R 2. In the comparison we use the same value of R 2 to compute critical values for each of the tests. In the rst column of the CADF results, where R 2 =0, we have essentially the same results as the Dickey and Fuller (1979) test in Table 2 that ignores the covariates. This should be the case, the included x t variable in the ADF regression has a population coecient of zero in this case. Likewise, the rst column of the ˆ(1; ) test matches with the P T test for the reasons we have described. This gives an insight into the dierence in the two approaches, the dierence between the CADF and (1; ) is similar to the dierence between the Dickey and Fuller (1979) approach and the ERS approach. When R 2 0, we see that the ˆ(1; ) test outperforms the CADF test in terms of power, although is slightly worse in size performance. The increases in power can be quite large. In the case 3 when R 2 =0:09 the power of the ˆ(1; ) test is 2 3 times that of the CADF test. For case 5 the eects are not as dramatic, but still power gains of 50% or so are available from using the covariates test proposed here over the CADF test.

13 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Unit root tests and long run structural VAR estimation Blanchard and Quah (1989) derive a method for identifying structural VAR s from restrictions placed on the spectral density of the data at frequency zero when there are known unit roots in the system. Consider the bivariate version of the model considered in this paper when we impose that the root is equal to unity, [ ] yt A(L) = t : x t Inverting the lag polynomial gives [ ] yt = C(L) t = C(L)KK 1 t = D(L) t ; x t where C(L)=A(L) 1 and E[ t t ]=I. This model is not identied in the usual sense as for any of the innite possible invertible matrices K we obtain a dierent structural model. In this bivariate system we require a single restriction so that the rotation K is unique for the model to be identied (this would be the order condition). In such systems, y t is permanently aected by shock(s) since it is an integrated process. On economic grounds, it may be interesting to identify the model such that only one of the structural shocks has a permanent eect on y t.inblanchard and Quah (1989) this argument meant that demand shocks could not have a permanent eect. In King et al. (1991) cointegration was used to imply a smaller number of permanent shocks than total shocks. In such cases it is possible to identify the model as the cumulated sum of the structural impulse responses, D(1), will be triangular as only one of the shocks has a long run eect on y t. For the model above, the identication scheme would set the (1,2) component of D(1) equal to zero. Since the spectral density of the data at frequency zero (scaled by 2) is = D(1)D(1), this amounts to taking the choleski decomposition of the estimated matrix ˆ. Such a restriction is only interesting and useful in identication when the o diagonals for are indeed nonzero, i.e. when R 2 0. The crux of this approach to identication clearly is that y t indeed does have a unit root. If instead there were no permanent eects then we would interpret D(1) dierently and would have no reason to make this matrix triangular. So in practice a useful hypothesis test to report in undertaking this method would be a test for a unit root in y t. Further, when the imposed restriction is indeed informative, then R 2 0 and hence we are exactly in the cases where the tests of this paper yield power gains over univariate testing. Typically, such tests for a unit root to provide evidence of the validity of this restriction are undertaken using Dickey Fuller (1979) tests (see Gali (1999) for example), which neither use the full information in the model nor are they the most powerful univariate tests. The tests derived in this paper provide a natural test of the basic identication assumption of the Blanchard and Quah identication scheme. We apply the tests derived here and other common tests to the Blanchard Quah data set. The data are quarterly data on income and unemployment for the US from 1950:2 to 1987:4, where unemployment is the stationary variable x t and income is the

14 88 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) y t variable. We include constants and time trends in both unemployment and income 7 (so the tests are from case 5) and follow Blanchard and Quah in choosing eight lags. The DF statistic is 1:78 and the DF-GLS statistic of ERS is 1:37. Neither is close to rejecting for a 5% or 10% test. The 5 (1; ) test is For the estimated R 2 of 0.76 the critical value is 16.56, so we have a p-value of 0.07 and fail (but only just) to reject 5% and so nd some support for the Blanchard and Quah assumption Conclusion Typically in economics correlation between the variables is the rule rather than the exception. Often these are implied by theory. Either way, this information can be extremely valuable in testing assumptions that are ancillary to the modeling process. This appears to be especially true in the case of testing for a unit root. Hansen (1995) showed this with tests he developed based around the statistic of Dickey and Fuller (1979). In a related paper Horvath and Watson (1995) showed that power gains are available when there are known cointegrating relationships (which are then stationary variables). We have shown here that even greater gains are possible. The statistics are simple to implement and yield extremely large gains in power when the covariates are relevant. The statistics we generate, useful in many areas, are directly applicable to testing the unit root assumption in the identication of structural VAR s from long run restrictions. These restrictions do not make sense unless there is a process with a unit root in the model, yet typically very low power tests are used to examine this assumption. The tests derived here will have much better power at detecting the mistaken use of this procedure. Acknowledgements The authors would like to thank the editor, referees and conference participants at the September 1999 CNMLE Conference at SvinklHv, the European Meeting of the Econometric Society (Winter 2001), UCSD, UCSC, University of Houston, FRB Atlanta, ASU, and the University of St Gallen for comments. References Blanchard, O., Quah, D., The dynamic eect of aggregate demand and aggregate supply disturbances. American Economic Review 79, Blanchard and Quah included a time trend in unemployment on the grounds that it was increasing over the sample. They had the equivalent of a time trend with a break for the oil shocks in income. We do not include a known break such as this, however not including the break if it were truly there (tests which search for such a break typically fail to reject the hypothesis of no break) biases us away from rejecting the unit root. 8 We do reject for 7 lags, but not for shorter lags than this.

15 G. Elliott, M. Jansson / Journal of Econometrics 115 (2003) Caporale, G.M., Pittis, N., Unit root testing using covariates: some theory and evidence. Oxford Bulletin of Economic Studies 61, Dickey, D.A., Fuller, W.A., Distribution of estimators for autoregressive time series with a unit root. Journal of the American Statistical Association 74, Elliott, G., Ecient tests for a unit root when the initial observation is drawn from its unconditional distribution. International Economic Review 40, Elliott, G., Jansson, M., Pesavento, E., Optimal power for testing potential cointegrating vectors with known parameters for nonstationary. manuscript, UCSD. Elliott, G., Rothenberg, T.J., Stock, J.H., Ecient tests for an autoregressive unit root. Econometrica 64, Gali, J., Technology, employment and the business cycle: do technology shocks explain aggregate uctuations?. American Economic Review 89, Hansen, B.E., Rethinking the univariate approach to unit root testing: using covariates to increase power. Econometric Theory 11, Horvath, M.T.K., Watson, M.W., Testing for cointegration when some of the cointegrating vectors are known. Econometric Theory 11, Johansen, S., Statistical analysis of cointegrating vectors. Journal of Economic Dynamics and Control 12, King, M.L., Robust tests for spherical symmetry and their application to least squares regression. Annals of Statistics 8, King, M.L., Towards a theory of point optimal testing. Econometric Reviews 6, King, R.G., Plosser, C.I., Stock, J.H., Watson, M.W., Stochastic trends and economic uctuations. American Economic Review 81,